mir_analyzer/

suppression.rs

//! Inline issue suppression via source comments.
//!
//! Lets users silence a single false positive without touching `mir.xml` or a
//! baseline file. A [`SuppressionMap`] is built once per file from its source
//! text and consulted as a final post-filter over the analyzer's issues
//! (`batch.rs`), so it applies uniformly across every emitting pass —
//! body analysis, the collector, class checks and dead-code detection.
//!
//! ## Recognised directives
//!
//! Native (preferred), matching the existing `@mir-check` convention:
//!
//! | Directive                  | Scope                                   |
//! |----------------------------|-----------------------------------------|
//! | `@mir-ignore [Kind …]`     | trailing comment → its line; otherwise the next code line |
//! | `@mir-ignore-line [Kind …]`      | the comment's own line            |
//! | `@mir-ignore-next-line [Kind …]` | the next physical line            |
//! | `@mir-ignore-file [Kind …]`      | the whole file                    |
//!
//! `@mir-suppress*` is accepted as an alias of `@mir-ignore*`.
//!
//! Third-party aliases for drop-in compatibility:
//!
//! | Directive                   | Scope / kinds                          |
//! |-----------------------------|----------------------------------------|
//! | `@psalm-suppress Kind …`    | like `@mir-ignore` (named kinds)       |
//! | `@suppress Kind …`          | like `@mir-ignore` (named kinds)       |
//! | `@phpstan-ignore-line`      | the comment's own line, all kinds      |
//! | `@phpstan-ignore-next-line` | the next line, all kinds               |
//! | `@phpstan-ignore …`         | the next line, all kinds               |
//!
//! When no `Kind` follows the directive, *all* issues on the target line are
//! suppressed. Kinds may be given by name (`UndefinedClass`) or by code
//! (`MIR0123`); multiple kinds are space- or comma-separated. PHPStan's
//! `@phpstan-ignore*` forms always suppress every kind on their target, since
//! PHPStan identifiers do not map onto mir's [`IssueKind`] names.
//!
//! [`IssueKind`]: mir_issues::IssueKind

use rustc_hash::{FxHashMap, FxHashSet};

/// Set of issue kinds a directive applies to.
#[derive(Debug, Clone)]
enum KindSet {
    /// Every kind on the target.
    All,
    /// Specific kinds, matched against `IssueKind::name()` or `code()`.
    Named(FxHashSet<String>),
}

impl KindSet {
    fn matches(&self, name: &str, code: &str) -> bool {
        match self {
            KindSet::All => true,
            // Kind names/codes are matched case-insensitively — a directive
            // author writing `@mir-ignore undefinedclass` (or any other casing
            // that doesn't exactly match `IssueKind::name()`'s PascalCase)
            // must still suppress the issue instead of silently doing nothing.
            // The set is small (a handful of kinds per directive at most), and
            // preserving each entry's original casing (rather than
            // lowercasing at parse time) keeps `UnusedSuppress`'s message
            // quoting the text the author actually wrote.
            KindSet::Named(set) => set.iter().any(|k| {
                k.eq_ignore_ascii_case(name)
                    || k.eq_ignore_ascii_case(code)
                    || alias_targets(k)
                        .iter()
                        .any(|t| t.eq_ignore_ascii_case(name))
            }),
        }
    }

    fn merge(&mut self, other: KindSet) {
        match (self, other) {
            // Already broadest possible.
            (KindSet::All, _) => {}
            (slot @ KindSet::Named(_), KindSet::All) => *slot = KindSet::All,
            (KindSet::Named(a), KindSet::Named(b)) => a.extend(b),
        }
    }
}

/// Third-party (Psalm/PHPStan) issue identifiers that don't match any
/// `IssueKind::name()` verbatim because mir models the same underlying
/// diagnostic under a different name. Without this table, a directive like
/// `@psalm-suppress PossiblyNullReference` neither suppresses the mir issue
/// it was meant to silence (`KindSet::matches` finds no name match) nor
/// counts as used (`unused_named`'s `kind_matches` finds no match either) —
/// so the real issue still fires AND the suppression itself gets flagged
/// `UnusedSuppress`, doubling the noise instead of silencing it.
///
/// Each entry maps one third-party name to every mir `IssueKind::name()` it
/// should be treated as matching. Looked up case-insensitively by both
/// `KindSet::matches` and `unused_named`, so it applies uniformly whether the
/// suppression ends up used for line-suppression or unused-detection.
const ISSUE_NAME_ALIASES: &[(&str, &[&str])] = &[
    // Psalm's `PossiblyNullReference` covers member access (property fetch OR
    // method call) on a possibly-null receiver; mir splits that into two
    // separate kinds by access form.
    (
        "PossiblyNullReference",
        &["PossiblyNullPropertyFetch", "PossiblyNullMethodCall"],
    ),
    // Psalm's check for a property that isn't definitely assigned by every
    // constructor path; mir's equivalent is named for the resulting state
    // rather than the constructor-focused cause.
    (
        "PropertyNotSetInConstructor",
        &["PropertyPossiblyUninitialized"],
    ),
];

/// Mir `IssueKind` names that `name` (a third-party identifier) should also
/// be treated as matching, per [`ISSUE_NAME_ALIASES`]. Empty if `name` isn't
/// a known third-party alias.
fn alias_targets(name: &str) -> &'static [&'static str] {
    ISSUE_NAME_ALIASES
        .iter()
        .find(|(alias, _)| alias.eq_ignore_ascii_case(name))
        .map_or(&[], |(_, targets)| *targets)
}

/// Where a directive applies, relative to the comment's own line.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum Scope {
    /// The comment's own physical line.
    SameLine,
    /// The next code line (next non-blank physical line).
    NextLine,
    /// Every line in the file.
    File,
}

struct Directive {
    scope: Scope,
    kinds: KindSet,
    /// For [`Scope::NextLine`]: whether to skip intervening comment lines (not
    /// just blanks) when locating the target. Set for "documents the following
    /// element" forms (`@psalm-suppress`, bare `@mir-ignore`, …) so a directive
    /// inside a multi-line docblock still lands on the declaration it annotates,
    /// past the closing `*/`.
    skip_comments: bool,
}

/// A named (non-All) suppression, tracked for `UnusedSuppress` detection.
/// `report_line` is where an `UnusedSuppress` issue is emitted if this
/// directive never matches anything; `covered_lines` is every line an
/// issue of `kind` may legitimately land on to count as "used" — normally
/// just `report_line` itself, but a multi-line signature's continuation
/// lines (where a per-parameter issue like `UnusedParam` actually fires)
/// are included too, since the directive covers those the same way
/// `SuppressionMap::lines` already does.
#[derive(Debug, Clone)]
pub struct NamedSuppression {
    pub report_line: u32,
    pub covered_lines: Vec<u32>,
    pub kind: String,
}

/// Per-file map of suppressed lines, built from source comments.
#[derive(Debug, Default)]
pub struct SuppressionMap {
    /// 1-based line number → kinds suppressed on that line.
    lines: FxHashMap<u32, KindSet>,
    /// Whole-file suppression, if any directive requested it.
    file: Option<KindSet>,
    /// Named (non-All) suppressions, for `UnusedSuppress` detection. Only
    /// `@psalm-suppress X` / `@suppress X` / `@mir-suppress X` forms populate
    /// this — blanket `@phpstan-ignore*` suppressions are intentionally excluded.
    pub named_suppressions: Vec<NamedSuppression>,
}

impl SuppressionMap {
    /// No directives — used to skip work for files with no suppression comments.
    pub fn is_empty(&self) -> bool {
        self.lines.is_empty() && self.file.is_none()
    }

    /// Whether an issue of `name`/`code` reported at 1-based `line` is suppressed.
    pub fn is_suppressed(&self, line: u32, name: &str, code: &str) -> bool {
        if let Some(file) = &self.file {
            if file.matches(name, code) {
                return true;
            }
        }
        self.lines.get(&line).is_some_and(|k| k.matches(name, code))
    }

    /// Scan `source` for suppression directives.
    pub fn from_source(source: &str) -> Self {
        if !contains_ascii_case(source, "ignore") && !contains_ascii_case(source, "suppress") {
            return SuppressionMap::default();
        }

        let raw_lines: Vec<&str> = source.lines().collect();
        let in_heredoc_body = heredoc_body_mask(&raw_lines);
        let block_comment_open = block_comment_open_mask(&raw_lines);
        let mut map = SuppressionMap::default();

        for (idx, raw) in raw_lines.iter().enumerate() {
            // A `#`/`//`-looking line INSIDE a heredoc/nowdoc body (embedded
            // shell/SQL/etc.) is not a real comment — without this, a line
            // like `# @mir-ignore-file UndefinedClass` in an embedded script
            // is indistinguishable from a genuine suppression directive.
            if in_heredoc_body[idx] {
                continue;
            }
            let Some((directive, track_named)) = parse_directive_with_tracking(raw) else {
                continue;
            };
            match directive.scope {
                Scope::File => match &mut map.file {
                    Some(existing) => existing.merge(directive.kinds),
                    None => map.file = Some(directive.kinds),
                },
                Scope::SameLine => {
                    let line_no = idx as u32 + 1;
                    if track_named {
                        if let KindSet::Named(ref names) = directive.kinds {
                            for name in names {
                                map.named_suppressions.push(NamedSuppression {
                                    report_line: line_no,
                                    covered_lines: vec![line_no],
                                    kind: name.clone(),
                                });
                            }
                        }
                    }
                    insert_line(&mut map.lines, line_no, directive.kinds);
                }
                Scope::NextLine => {
                    let (target, extra_covered_lines) = next_code_line(
                        &raw_lines,
                        idx,
                        directive.skip_comments,
                        block_comment_open[idx],
                    );
                    if track_named {
                        if let KindSet::Named(ref names) = directive.kinds {
                            // A continuation line of `target`'s own multi-line
                            // signature (a later parameter, e.g.) is where a
                            // per-parameter issue like `UnusedParam` actually
                            // fires — track it alongside `target` as a single
                            // suppression that's "used" if EITHER line has a
                            // matching issue, so a genuinely-used directive
                            // whose issue lands on a continuation line isn't
                            // misreported as unused.
                            let mut covered_lines = extra_covered_lines.clone();
                            covered_lines.push(target);
                            for name in names {
                                map.named_suppressions.push(NamedSuppression {
                                    report_line: target,
                                    covered_lines: covered_lines.clone(),
                                    kind: name.clone(),
                                });
                            }
                        }
                    }
                    // An attribute line skipped en route to `target`, or a
                    // continuation line of `target`'s own multi-line
                    // signature, stays covered by the same directive — an
                    // issue about the attribute itself (e.g.
                    // `UndefinedAttributeClass`) or about a later parameter
                    // (e.g. `UnusedParam`) fires there, not at `target`.
                    for extra_line in extra_covered_lines {
                        insert_line(&mut map.lines, extra_line, directive.kinds.clone());
                    }
                    insert_line(&mut map.lines, target, directive.kinds);
                }
            }
        }

        map
    }

    /// Returns unused named suppressions: those that did not match any issue
    /// in `all_issues`. The returned vec contains `(target_line, kind_name)`.
    ///
    /// `pre_suppressed` is the subset of `all_issues` that arrived already
    /// suppressed (via the `IssueBuffer` mechanism in collector/body analysis).
    /// These may be emitted at a different line than the suppression target
    /// (e.g. `InvalidDocblock` at a docblock-start line vs. the following
    /// declaration line), so they are matched within a 30-line window before
    /// the target.
    pub fn unused_named(
        &self,
        all_issues: &[&mir_issues::Issue],
        pre_suppressed: &[&mir_issues::Issue],
    ) -> Vec<(u32, String)> {
        self.named_suppressions
            .iter()
            .filter(|ns| {
                let target_line = ns.report_line;
                let kind = &ns.kind;
                // Compare case-insensitively, same as `KindSet::matches` — a
                // directive can name a kind in any casing, and may also name
                // a third-party (Psalm/PHPStan) identifier for a check mir
                // models under a different name (`alias_targets`).
                let kind_matches = |issue: &&mir_issues::Issue| {
                    let display_name = issue.kind.display_name();
                    display_name.eq_ignore_ascii_case(kind.as_str())
                        || issue.kind.code().eq_ignore_ascii_case(kind.as_str())
                        || alias_targets(kind.as_str())
                            .iter()
                            .any(|t| t.eq_ignore_ascii_case(display_name))
                };
                // Normal case: a SuppressionMap-suppressed issue at the report
                // line itself, OR any other line this same directive covers
                // (a multi-line signature's continuation lines, where a
                // per-parameter issue like `UnusedParam` actually fires).
                let at_covered_line = all_issues.iter().any(|issue| {
                    ns.covered_lines.contains(&issue.location.line) && kind_matches(issue)
                });
                if at_covered_line {
                    return false; // suppression IS used
                }
                // Docblock case: collector-emitted issues (like `InvalidDocblock`)
                // land at the docblock-start line, which precedes the declaration
                // that the suppression targets. Allow a 100-line look-back so a
                // `@psalm-suppress InvalidDocblock` in an unusually long
                // multi-line docblock (heavy @template/@method/@property lists)
                // is still recognised as used even though its issue line !=
                // target_line — 30 lines was too easy to exceed for such a
                // docblock, wrongly reporting a real, used suppression as unused.
                let min_line = target_line.saturating_sub(100);
                let covered_by_pre_suppressed = pre_suppressed.iter().any(|issue| {
                    issue.location.line >= min_line
                        && issue.location.line < target_line
                        && kind_matches(issue)
                });
                !covered_by_pre_suppressed
            })
            .map(|ns| (ns.report_line, ns.kind.clone()))
            .collect()
    }
}

fn contains_ascii_case(haystack: &str, needle: &str) -> bool {
    haystack
        .as_bytes()
        .windows(needle.len())
        .any(|window| window.eq_ignore_ascii_case(needle.as_bytes()))
}

fn insert_line(lines: &mut FxHashMap<u32, KindSet>, line: u32, kinds: KindSet) {
    match lines.get_mut(&line) {
        Some(existing) => existing.merge(kinds),
        None => {
            lines.insert(line, kinds);
        }
    }
}

/// Per-line mask: `true` for a physical line that falls INSIDE a heredoc or
/// nowdoc body (between the `<<<IDENT` opener line and its closing `IDENT`
/// line), `false` everywhere else — including the opener line itself, which
/// still carries real PHP code (`$x = <<<EOT`) before the body starts.
///
/// Deliberately simple, matching this file's existing line-based scanning:
/// doesn't defend against a `<<<` that itself appears inside an unrelated
/// string literal on the same line (a separate, rarer gap) — only real
/// heredoc/nowdoc openers are expected to actually appear this way in
/// practice.
fn heredoc_body_mask(raw_lines: &[&str]) -> Vec<bool> {
    let mut mask = vec![false; raw_lines.len()];
    let mut closing_ident: Option<&str> = None;
    for (idx, line) in raw_lines.iter().enumerate() {
        if let Some(ident) = closing_ident {
            mask[idx] = true;
            if is_heredoc_closing_line(line, ident) {
                closing_ident = None;
            }
            continue;
        }
        closing_ident = find_heredoc_opener(line);
    }
    mask
}

/// If `line` opens a heredoc (`<<<IDENT`) or nowdoc (`<<<'IDENT'`/`<<<"IDENT"`),
/// return the closing identifier to watch for.
fn find_heredoc_opener(line: &str) -> Option<&str> {
    let pos = line.find("<<<")?;
    let rest = line[pos + 3..].trim_start();
    let (quote, rest) = match rest.as_bytes().first() {
        Some(b'\'') => (Some(b'\''), &rest[1..]),
        Some(b'"') => (Some(b'"'), &rest[1..]),
        _ => (None, rest),
    };
    let end = rest
        .find(|c: char| !c.is_ascii_alphanumeric() && c != '_')
        .unwrap_or(rest.len());
    let ident = &rest[..end];
    if ident.is_empty() {
        return None;
    }
    if let Some(q) = quote {
        if rest.as_bytes().get(end) != Some(&q) {
            return None;
        }
    }
    Some(ident)
}

/// Whether `line` is the closing line of a heredoc/nowdoc body started with
/// `ident` — optional leading whitespace (PHP 7.3+ flexible heredoc
/// indentation), then the identifier as a whole word (not a prefix of a
/// longer identifier).
fn is_heredoc_closing_line(line: &str, ident: &str) -> bool {
    let Some(rest) = line.trim_start().strip_prefix(ident) else {
        return false;
    };
    rest.chars()
        .next()
        .is_none_or(|c| !c.is_ascii_alphanumeric() && c != '_')
}

/// Per-line mask: `true` when a `/* … */` block comment (most commonly a
/// docblock) is still open at the END of that physical line — its closing
/// `*/` hasn't been seen yet. Lets a `NextLine` directive found on one of a
/// docblock's own continuation lines tell whether the lines immediately
/// following it are still the SAME comment (further tag lines, then its own
/// closing `*/`) rather than real code or a separate, later comment.
///
/// Deliberately simple, matching this file's other line-based scanners: no
/// quote-awareness, so a `/*`/`*/`-looking substring inside a string literal
/// is out of scope.
fn block_comment_open_mask(raw_lines: &[&str]) -> Vec<bool> {
    let mut mask = vec![false; raw_lines.len()];
    let mut open = false;
    for (idx, line) in raw_lines.iter().enumerate() {
        let bytes = line.as_bytes();
        let mut i = 0;
        while i < bytes.len() {
            if open {
                if bytes[i] == b'*' && bytes.get(i + 1) == Some(&b'/') {
                    open = false;
                    i += 2;
                } else {
                    i += 1;
                }
            } else if bytes[i] == b'/' && bytes.get(i + 1) == Some(&b'*') {
                open = true;
                i += 2;
            } else {
                i += 1;
            }
        }
        mask[idx] = open;
    }
    mask
}

/// Locate a directive's target line strictly after `idx`, as a 1-based
/// number, plus every single-line PHP 8 attribute (`#[Foo]`) line skipped
/// along the way.
///
/// Always skips blank lines and an attribute line sitting between the
/// directive and the declaration it modifies — real code, but never what a
/// suppression directive means to target (the declaration that follows).
/// The skipped attribute lines are returned too, not just discarded: an
/// issue about the attribute ITSELF (e.g. `UndefinedAttributeClass`, "this
/// attribute class does not exist") fires on the attribute's own line, not
/// the final target, and must stay covered by the same directive as the
/// declaration it annotates. When `skip_comments` is set, also skips
/// comment-only lines (`//`, `#`, `/* … */`, ` * …` docblock bodies and the
/// closing `*/`) so a directive written inside a multi-line docblock lands
/// on the declaration that follows it. Falls back to `idx + 2` when nothing
/// qualifies, so the directive still has a deterministic target.
///
/// `in_open_block_comment` is whether the directive's OWN line sits inside a
/// still-unterminated `/* … */` block comment — a multi-line docblock whose
/// closing `*/` hasn't been seen yet. Those remaining lines (further tag
/// lines, then the closing `*/`) can never be the target regardless of
/// `skip_comments`, since they're still the directive's own comment, not a
/// separate one — `skip_comments` only decides whether to also hunt past a
/// DIFFERENT, later comment sitting between that closing `*/` and the
/// declaration.
fn next_code_line(
    raw_lines: &[&str],
    idx: usize,
    skip_comments: bool,
    in_open_block_comment: bool,
) -> (u32, Vec<u32>) {
    let mut extra_covered_lines = Vec::new();
    // Net `[`/`]` depth of an in-progress multi-line attribute (`#[` opener
    // whose closing `]` is on a later line) — 0 when not currently inside
    // one. Every line consumed while this is positive is itself an
    // attribute line, same treatment the single-line case already gets.
    let mut multiline_attr_depth: i32 = 0;
    let mut in_open_block_comment = in_open_block_comment;
    for (offset, line) in raw_lines.iter().enumerate().skip(idx + 1) {
        let trimmed = line.trim();
        if in_open_block_comment {
            if trimmed.contains("*/") {
                in_open_block_comment = false;
            }
            continue;
        }
        if multiline_attr_depth > 0 {
            extra_covered_lines.push(offset as u32 + 1);
            multiline_attr_depth += bracket_delta(trimmed);
            continue;
        }
        if trimmed.is_empty() {
            continue;
        }
        if is_attribute_only(trimmed) {
            extra_covered_lines.push(offset as u32 + 1);
            continue;
        }
        // A `#[` opener with no matching `]` on the SAME line (net positive
        // bracket delta) starts a multi-line attribute — track depth across
        // subsequent lines instead of wrongly treating this opener line as
        // the real target.
        if trimmed.starts_with("#[") {
            let delta = bracket_delta(trimmed);
            if delta > 0 {
                extra_covered_lines.push(offset as u32 + 1);
                multiline_attr_depth = delta;
                continue;
            }
        }
        if skip_comments && is_comment_only(trimmed) {
            continue;
        }
        // A declaration whose own signature spans multiple physical lines
        // (one parameter per line is a common style) still has exactly one
        // target line for suppression purposes, yet per-line issues land
        // later in the SAME signature (e.g. `UnusedParam` on a parameter
        // several lines down). Track paren depth from the target line and
        // cover every continuation line until the signature's parens close.
        let target = offset as u32 + 1;
        let mut paren_depth = paren_delta(trimmed);
        let mut cont_idx = offset + 1;
        while paren_depth > 0 {
            let Some(cont_line) = raw_lines.get(cont_idx) else {
                break;
            };
            extra_covered_lines.push(cont_idx as u32 + 1);
            paren_depth += paren_delta(cont_line.trim());
            cont_idx += 1;
        }
        // A "documents the following element" directive above a function or
        // method covers its whole body in real Psalm/PHPStan semantics, not
        // just the signature line — extend coverage to every line spanned by
        // the function's own `{ ... }`, if it has one (abstract/interface
        // methods don't).
        if skip_comments && is_function_like(trimmed) {
            if let Some(body_lines) = function_body_lines(raw_lines, offset) {
                extra_covered_lines.extend(body_lines);
            }
        }
        return (target, extra_covered_lines);
    }
    (idx as u32 + 2, extra_covered_lines)
}

/// Whether a trimmed line is itself a function/method DECLARATION (after
/// skipping any leading visibility/`static`/`abstract`/`final` modifiers, the
/// next word is `function`). Used to decide whether a "documents the
/// following element" suppression directive's target has a body worth
/// extending coverage into.
///
/// Deliberately does not match on `function` appearing anywhere in the line —
/// a directive whose target statement merely CONTAINS a closure (`return
/// function () {…};`, `$cb = function () {…};`, an `array_map(function ()
/// {…}, …)` call) annotates that statement, not the closure's own nested
/// scope, so it must not be treated as covering the closure's body.
fn is_function_like(trimmed: &str) -> bool {
    const MODIFIERS: &[&str] = &[
        "public",
        "private",
        "protected",
        "static",
        "abstract",
        "final",
    ];
    for word in trimmed.split_whitespace() {
        let lower = word.to_ascii_lowercase();
        if MODIFIERS.contains(&lower.as_str()) {
            continue;
        }
        return lower == "function" || lower.starts_with("function(");
    }
    false
}

/// Physical line numbers (1-based, excluding `start_idx`'s own line) spanned
/// by a function/method body starting at 0-based `start_idx`, found by
/// scanning forward for a balanced `{`...`}` pair. Returns `None` if a bare
/// `;` terminates the declaration before any `{` appears — an abstract or
/// interface method, which has no body to extend suppression coverage into.
/// No quote-awareness, same conservative scope as `paren_delta`/`bracket_delta`:
/// a stray unmatched `{`/`}` inside a plain string literal (as opposed to
/// `"{$expr}"` interpolation, which is always balanced) could throw off depth
/// tracking — an accepted rare edge case.
fn function_body_lines(raw_lines: &[&str], start_idx: usize) -> Option<Vec<u32>> {
    let mut depth: i32 = 0;
    let mut opened = false;
    let mut covered = Vec::new();
    for (offset, line) in raw_lines.iter().enumerate().skip(start_idx) {
        for c in line.chars() {
            match c {
                '{' => {
                    depth += 1;
                    opened = true;
                }
                '}' => depth -= 1,
                ';' if !opened && depth == 0 => return None,
                _ => {}
            }
        }
        if offset != start_idx {
            covered.push(offset as u32 + 1);
        }
        if opened && depth <= 0 {
            return Some(covered);
        }
    }
    None
}

/// Net count of `(` minus `)` in `s` — no quote-awareness, matching
/// `bracket_delta`'s same conservative scope.
fn paren_delta(s: &str) -> i32 {
    s.chars().fold(0i32, |depth, c| match c {
        '(' => depth + 1,
        ')' => depth - 1,
        _ => depth,
    })
}

/// Net count of `[` minus `]` in `s` — no quote-awareness, matching the
/// same conservative scope `is_attribute_only`'s plain `ends_with(']')`
/// already accepts (real attribute argument text containing a bracket
/// inside a string literal is out of scope here).
fn bracket_delta(s: &str) -> i32 {
    s.chars().fold(0i32, |depth, c| match c {
        '[' => depth + 1,
        ']' => depth - 1,
        _ => depth,
    })
}

/// Whether a trimmed line is purely a comment (no PHP code). `#[` is treated as
/// a PHP 8 attribute (code), not a `#` comment.
fn is_comment_only(trimmed: &str) -> bool {
    trimmed.starts_with("//")
        || trimmed.starts_with("/*")
        || trimmed.starts_with('*')
        || (trimmed.starts_with('#') && !trimmed.starts_with("#["))
}

/// A single-line PHP 8 attribute (`#[Foo]`, `#[Foo(bar: 1)]`) with no other
/// code on the same line — a trailing same-line comment (`#[Foo] // note`,
/// `#[Foo] /* note */`) is stripped first, so `]` needn't end the line.
/// Doesn't attempt multi-line
/// bracket-depth tracking for a `#[` that spans several lines — only the
/// common single-line case, mirroring the scope this suppression logic
/// already accepts elsewhere (e.g. no nested-comment tracking either).
fn is_attribute_only(trimmed: &str) -> bool {
    if !trimmed.starts_with("#[") {
        return false;
    }
    // Skip the leading `#` (part of the attribute marker, not a comment)
    // before scanning for a trailing comment introducer — the same
    // string-literal-aware scan `extract_comment` uses for a suppression
    // directive's own line, applied here to the attribute's line instead.
    let rest = &trimmed[1..];
    let body = match find_comment_introducer(rest) {
        Some(pos) => rest[..pos].trim_end(),
        None => rest,
    };
    body.ends_with(']')
}

/// Directive keyword table, ordered longest-first so that, e.g.,
/// `@mir-ignore-next-line` is matched before the `@mir-ignore` prefix.
///
/// Each entry is `(keyword, scope, force_all)`. `force_all` makes the directive
/// suppress every kind regardless of trailing tokens (PHPStan semantics).
const KEYWORDS: &[(&str, Scope, bool)] = &[
    ("@mir-ignore-next-line", Scope::NextLine, false),
    ("@mir-suppress-next-line", Scope::NextLine, false),
    ("@phpstan-ignore-next-line", Scope::NextLine, true),
    ("@mir-ignore-line", Scope::SameLine, false),
    ("@mir-suppress-line", Scope::SameLine, false),
    ("@phpstan-ignore-line", Scope::SameLine, true),
    ("@mir-ignore-file", Scope::File, false),
    ("@mir-suppress-file", Scope::File, false),
    // Bare forms (scope resolved below from comment position).
    ("@mir-ignore", Scope::NextLine, false),
    ("@mir-suppress", Scope::NextLine, false),
    ("@psalm-suppress", Scope::NextLine, false),
    ("@suppress", Scope::NextLine, false),
    ("@phpstan-ignore", Scope::NextLine, true),
];

/// Bare directives (no `-line`/`-next-line`/`-file` suffix) resolve their scope
/// from where the comment sits: a trailing comment annotates its own line, a
/// standalone comment annotates the statement that follows it.
const BARE_KEYWORDS: &[&str] = &[
    "@mir-ignore",
    "@mir-suppress",
    "@psalm-suppress",
    "@suppress",
    "@phpstan-ignore",
];

/// Like `parse_directive` (which is parse_directive_with_tracking discarding the tracking flag),
/// but also returns whether named suppression tracking
/// should be applied (true for `@psalm-suppress`, `@mir-suppress`, `@suppress`
/// and `@mir-ignore` forms; false for `@phpstan-*` which are blanket suppressors
/// not tied to specific issue kinds).
fn parse_directive_with_tracking(raw: &str) -> Option<(Directive, bool)> {
    let comment = extract_comment(raw)?;

    // A comment can carry more than one directive-shaped substring (e.g. a
    // user's own `@mir-ignore-line Foo` followed later by an unrelated
    // `@phpstan-ignore-next-line`) — pick whichever keyword actually appears
    // FIRST in the text, not the first entry `KEYWORDS` happens to define,
    // so an earlier table entry never shadows a later-defined keyword that
    // the author actually wrote earlier in the comment.
    let mut best: Option<(usize, &str, Scope, bool)> = None;
    for &(keyword, scope, force_all) in KEYWORDS {
        let Some(pos) = find_ci(comment.content, keyword) else {
            continue;
        };
        // Reject keyword matches that are really a prefix of a longer token
        // (e.g. `@mir-ignore` inside `@mir-ignore-line`).
        let after = &comment.content[pos + keyword.len()..];
        if after
            .chars()
            .next()
            .is_some_and(|c| c.is_ascii_alphanumeric() || c == '-')
        {
            continue;
        }
        let is_leftmost = match best {
            None => true,
            Some((best_pos, ..)) => pos < best_pos,
        };
        if is_leftmost {
            best = Some((pos, keyword, scope, force_all));
        }
    }
    let (pos, keyword, scope, force_all) = best?;
    let after = &comment.content[pos + keyword.len()..];

    let is_bare = BARE_KEYWORDS.contains(&keyword);

    // Bare forms: a trailing comment suppresses its own line.
    let scope = if is_bare && comment.has_code_before {
        Scope::SameLine
    } else {
        scope
    };

    // The "documents the following element" forms (bare `@psalm-suppress`,
    // `@mir-ignore`, …) also skip past a SEPARATE comment sitting between the
    // directive's own comment and the declaration. PHPStan's explicit
    // `*-next-line` and bare `@phpstan-ignore` keep literal
    // next-non-blank-line semantics there — but the directive's own
    // enclosing block comment (a multi-line docblock's remaining tag lines
    // and closing `*/`) is always exited first regardless, via
    // `next_code_line`'s `in_open_block_comment` handling, since a line still
    // inside that comment can never be the target either way.
    let skip_comments = scope == Scope::NextLine && is_bare && !force_all;

    let kinds = if force_all {
        KindSet::All
    } else {
        parse_kinds(after)
    };

    // Track named suppressions only for non-phpstan forms (phpstan forms
    // always suppress all kinds, so they can never be "unused for a specific kind").
    let track_named = !keyword.starts_with("@phpstan");

    Some((
        Directive {
            scope,
            kinds,
            skip_comments,
        },
        track_named,
    ))
}

/// Case-insensitive `str::find` for an ASCII directive keyword. Kind names
/// are already matched case-insensitively (`KindSet::matches`); the keyword
/// itself was still exact-case, so `@MIR-IGNORE`/`@Psalm-Suppress` silently
/// matched nothing. `to_ascii_lowercase` never changes a string's byte
/// length (only ASCII bytes are touched), so the returned index stays valid
/// against the original, non-lowercased `haystack`.
fn find_ci(haystack: &str, needle: &str) -> Option<usize> {
    haystack
        .to_ascii_lowercase()
        .find(&needle.to_ascii_lowercase())
}

struct Comment<'a> {
    /// Text from the comment introducer onward (still includes `*/`, `*`, etc.).
    content: &'a str,
    /// Whether non-whitespace code precedes the comment on the same line.
    has_code_before: bool,
}

/// Isolate the comment portion of a physical line, if any. Handles `//`, `#`
/// and `/* … */` introducers, block-comment continuation lines (` * …`) and
/// bare directive lines inside block comments (`@psalm-suppress …`).
fn extract_comment(raw: &str) -> Option<Comment<'_>> {
    let trimmed = raw.trim_start();

    // Block-comment continuation or a bare directive line: no code precedes it.
    if trimmed.starts_with('*') {
        return Some(Comment {
            content: trimmed.trim_start_matches('*'),
            has_code_before: false,
        });
    }
    if trimmed.starts_with('@') {
        return Some(Comment {
            content: trimmed,
            has_code_before: false,
        });
    }

    // Earliest single-line / block introducer on the line, skipping any
    // that fall inside a quoted string literal (`$x = "// not a
    // comment";` carries no comment). Only a same-line string is handled
    // (a bounded, quote-toggle scan); a heredoc/multi-line string body
    // needs cross-line state this line-at-a-time scanner doesn't have.
    let pos = find_comment_introducer(raw)?;
    let has_code_before = !raw[..pos].trim().is_empty();
    Some(Comment {
        content: &raw[pos..],
        has_code_before,
    })
}

/// Byte offset of the earliest `//`, `#`, or `/*` comment introducer in
/// `raw`, ignoring any that fall inside a single- or double-quoted string
/// literal (with backslash-escape awareness). Returns `None` if the line
/// has no real comment introducer at all.
fn find_comment_introducer(raw: &str) -> Option<usize> {
    let bytes = raw.as_bytes();
    let mut in_squote = false;
    let mut in_dquote = false;
    let mut i = 0;
    while i < bytes.len() {
        let c = bytes[i];
        if in_squote {
            i += if c == b'\\' && i + 1 < bytes.len() {
                2
            } else {
                if c == b'\'' {
                    in_squote = false;
                }
                1
            };
            continue;
        }
        if in_dquote {
            i += if c == b'\\' && i + 1 < bytes.len() {
                2
            } else {
                if c == b'"' {
                    in_dquote = false;
                }
                1
            };
            continue;
        }
        match c {
            b'\'' => in_squote = true,
            b'"' => in_dquote = true,
            b'/' if bytes.get(i + 1) == Some(&b'/') || bytes.get(i + 1) == Some(&b'*') => {
                return Some(i);
            }
            // `#[` is a PHP 8 attribute opener, not a `#` line comment (see
            // `is_comment_only`'s identical carve-out) — keep scanning past
            // it instead of stopping here, so a REAL trailing comment after
            // the attribute (`#[Attr] // note`) is still found.
            b'#' if bytes.get(i + 1) == Some(&b'[') => {}
            b'#' => return Some(i),
            _ => {}
        }
        i += 1;
    }
    None
}

/// Collect issue kind names/codes following a directive keyword. Kinds are
/// comma-separated (`Foo, Bar`); only the FIRST whitespace-delimited word of
/// each comma-separated segment can be a kind name — every further word in
/// that same segment is treated as free-text prose (e.g. a trailing
/// explanation like `UndefinedClass because of a vendor stub`), not another
/// kind, since a real multi-kind list always uses a comma to continue. Once a
/// segment has trailing words after its kind name, everything from there on
/// — including later comma-separated segments — is that same free-text
/// prose, so a comma inside the prose (`Foo because of X, not fully typed`)
/// must not be mistaken for the start of another kind. Stops entirely at the
/// block-comment terminator. An empty result means "all kinds".
fn parse_kinds(rest: &str) -> KindSet {
    let mut set = FxHashSet::default();
    'segments: for segment in rest.split(',') {
        let mut tokens = segment
            .split([' ', '\t'])
            .map(str::trim)
            .filter(|t| !t.is_empty());
        for token in tokens.by_ref() {
            // End of the comment / docblock — stop scanning entirely.
            if token.starts_with("*/") || token.starts_with('*') {
                break 'segments;
            }
            // A kind name is alphanumeric (plus `_`); anything else is
            // skipped, keeping the next token as the segment's candidate.
            if token.chars().all(|c| c.is_ascii_alphanumeric() || c == '_') {
                set.insert(token.to_string());
                if tokens.next().is_some() {
                    // Trailing prose follows in this segment — the rest of
                    // the comment, comma or not, is that same free text.
                    break 'segments;
                }
                continue 'segments;
            }
        }
    }
    if set.is_empty() {
        KindSet::All
    } else {
        KindSet::Named(set)
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    fn map(src: &str) -> SuppressionMap {
        SuppressionMap::from_source(src)
    }

    #[test]
    fn line_comment_above_statement_suppresses_next_line() {
        // line 2 comment → suppress line 3
        let m = map("<?php\n// @psalm-suppress UndefinedClass\nnew NoSuchClass();\n");
        assert!(m.is_suppressed(3, "UndefinedClass", "MIR0000"));
        assert!(!m.is_suppressed(2, "UndefinedClass", "MIR0000"));
    }

    #[test]
    fn trailing_comment_suppresses_own_line() {
        let m = map("<?php\nnew NoSuchClass(); // @mir-ignore UndefinedClass\n");
        assert!(m.is_suppressed(2, "UndefinedClass", "MIR0000"));
    }

    #[test]
    fn single_line_docblock_above_statement() {
        let m = map("<?php\n/** @psalm-suppress UndefinedClass */\nnew NoSuchClass();\n");
        assert!(m.is_suppressed(3, "UndefinedClass", "MIR0000"));
    }

    #[test]
    fn phpstan_ignore_next_line_suppresses_all() {
        let m = map("<?php\n// @phpstan-ignore-next-line\nnew NoSuchClass();\n");
        assert!(m.is_suppressed(3, "UndefinedClass", "MIR0000"));
        assert!(m.is_suppressed(3, "AnyOtherKind", "MIR9999"));
    }

    #[test]
    fn ignore_line_targets_own_line() {
        let m = map("<?php\nnew NoSuchClass(); // @mir-ignore-line\n");
        assert!(m.is_suppressed(2, "UndefinedClass", "MIR0000"));
    }

    #[test]
    fn next_line_skips_blank_lines() {
        let m = map("<?php\n/** @psalm-suppress UndefinedClass */\n\n\nnew NoSuchClass();\n");
        assert!(m.is_suppressed(5, "UndefinedClass", "MIR0000"));
    }

    #[test]
    fn multiline_docblock_skips_to_declaration() {
        // line 2: /**, line 3: * @psalm-suppress, line 4: */, line 5: declaration.
        let src =
            "<?php\n/**\n * @psalm-suppress UnusedMethod\n */\nprivate function a(): void {}\n";
        let m = map(src);
        assert!(m.is_suppressed(5, "UnusedMethod", "MIR0000"));
    }

    #[test]
    fn phpstan_next_line_is_literal_not_comment_skipping() {
        // PHPStan's -next-line targets the next non-blank line even if it's a
        // comment; it does not hunt for the next code line.
        let m = map("<?php\n// @phpstan-ignore-next-line\n// unrelated comment\nfoo();\n");
        assert!(m.is_suppressed(3, "X", "MIR0000"));
        assert!(!m.is_suppressed(4, "X", "MIR0000"));
    }

    #[test]
    fn named_kind_does_not_suppress_other_kinds() {
        let m = map("<?php\n// @mir-ignore UndefinedClass\nfoo();\n");
        assert!(m.is_suppressed(3, "UndefinedClass", "MIR0000"));
        assert!(!m.is_suppressed(3, "UndefinedFunction", "MIR0001"));
    }

    #[test]
    fn match_by_code() {
        let m = map("<?php\n// @mir-ignore MIR1400\nfoo();\n");
        assert!(m.is_suppressed(3, "ParseError", "MIR1400"));
    }

    #[test]
    fn file_scope_suppresses_every_line() {
        let m = map("<?php // @mir-ignore-file UndefinedClass\nfoo();\nbar();\n");
        assert!(m.is_suppressed(2, "UndefinedClass", "MIR0000"));
        assert!(m.is_suppressed(99, "UndefinedClass", "MIR0000"));
        assert!(!m.is_suppressed(2, "UndefinedFunction", "MIR0001"));
    }

    #[test]
    fn multiple_kinds_one_directive() {
        let m = map("<?php\n// @psalm-suppress UndefinedClass, NullMethodCall\nfoo();\n");
        assert!(m.is_suppressed(3, "UndefinedClass", "MIR0000"));
        assert!(m.is_suppressed(3, "NullMethodCall", "MIR0001"));
    }

    #[test]
    fn no_directive_is_empty() {
        let m = map("<?php\n$x = \"@psalm-suppress not a comment\";\nfoo();\n");
        // It's inside a string but after `//`? No `//` here, so not detected.
        assert!(m.is_empty());
    }

    #[test]
    fn prefix_is_not_confused_with_longer_keyword() {
        // `@mir-ignore-next-line` must be parsed as next-line, not bare same-line.
        let m = map("<?php\nfoo(); // @mir-ignore-next-line\nbar();\n");
        assert!(m.is_suppressed(3, "AnyKind", "MIR0000"));
        assert!(!m.is_suppressed(2, "AnyKind", "MIR0000"));
    }
}